Re: [PATCH] gpio: Allow user to customise maximum number of GPIOs

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Thu, Aug 18, 2022 at 1:33 PM Arnd Bergmann <arnd@xxxxxxxx> wrote:
> On Thu, Aug 18, 2022 at 1:13 PM Linus Walleij <linus.walleij@xxxxxxxxxx> wrote:

> > static inline bool gpio_is_valid(int number)
> > {
> >         return number >= 0 && number < ARCH_NR_GPIOS;
> > }
> >
> > ?
> >
> > If using GPIO descriptors, any descriptor != NULL is valid,
> > this one is just used with legacy GPIOs. Maybe we should just
> > delete gpio_is_valid() everywhere and then drop the cap?
>
> I think it makes sense to keep gpio_is_valid() for as long as we
> support the numbers.

Hmmm....

> > I think there may be systems and users that still depend on GPIO base
> > numbers being assigned from ARCH_NR_GPIOS and
> > downwards (userspace GPIO numbers in sysfs will also change...)
> > otherwise we could assign from 0 and up.
>
> Is it possible to find in-kernel users that depend on well-known
> numbers for dynamically assigned gpios? I would argue
> that those are always broken.

Most in-kernel users hard-code the base to something like
0 etc it's only the ones that code -1 into .base that are in
trouble because that will activate dynamic assignment for the
base.

git grep 'base = -1' yields these suspects:

arch/arm/common/sa1111.c:       sachip->gc.base = -1;
arch/arm/common/scoop.c:        devptr->gpio.base = -1;
arch/powerpc/platforms/52xx/mpc52xx_gpt.c:      gpt->gc.base = -1;
arch/powerpc/platforms/83xx/mcu_mpc8349emitx.c: gc->base = -1;

That's all! We could just calculate these to 512-ngpios and
hardcode that instead.

> Even for the sysfs interface, it is questionable to rely on
> specific numbers because at least in an arm multiplatform
> kernel the top number changes based on kernel configuration.

Yeah :/ still these users tend to angrily report any breakage
due to expected (fragile) behaviour.

> > Right now the safest would be:
> > Assign from 512 and downwards until we hit 0 then assign
> > from something high, like U32_MAX and downward.
> >
> > That requires dropping gpio_is_valid() everywhere.
> >
> > If we wanna be bold, just delete gpio_is_valid() and assign
> > bases from 0 and see what happens. But I think that will
> > lead to regressions.
>
> I'm still unsure how removing gpio_is_valid() would help.

If we allow GPIO base all the way to U32_MAX
this function becomes:

static inline bool gpio_is_valid(int number)
{
        return number >= 0 && number < U32_MAX;
}

and we can then just

#define gpio_is_valid true

and in that case it is better to delete the use of this function
altogether since it can not fail.

> What I could imagine as a next step would be to mark all
> consumer drivers and the sysfs interface that use gpio
> numbers as 'depends on GPIO_LEGACY' and then only
> provide the corresponding drivers if that option is set.

Hm I wonder what Bartosz and Alexandre Courbot and thinks
about a GPIO_LEGACY symbol to phase out the global
GPIO numberspace. I kind of like the idea.

I made the sysfs depend on CONFIG_EXPERT to at least make it less
accessible and not provide users with guns to shoot themselves
in the foot.

Yours,
Linus Walleij



[Index of Archives]     [Kernel Newbies]     [Security]     [Netfilter]     [Bugtraq]     [Linux FS]     [Yosemite Forum]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Samba]     [Video 4 Linux]     [Device Mapper]     [Linux Resources]

  Powered by Linux